Pinned
Chen Tessler
1,885 posts
Chen Tessler
@ChenTessler
Research Scientist @NVIDIAAI ; Training humanoids ๐ค ; Reinforcement learning PhD @TechnionLive ๐ฎ๐ฑ. Views are my own.
- Excited to share our latest work! ๐คฉ Masked Mimic ๐ฅท: Unified Physics-Based Character Control Through Masked Motion Inpainting Project page: research.nvidia.com/labs/par/maskeโฆ with: Yunrong (Kelly) Guo, @ofirnabati, @GalChechik and @xbpeng4. @SIGGRAPHAsia (ACM TOG). 1/ Read
- Replying to @hardmaruDon't forget to teach them some Bayesian probability! They should have good control over the basics before they jump into deep learning voodoo!
- ๐ฅณ๐ฅณ๐ฅณ SMPL Motion Tracker Released! ๐ฅณ๐ฅณ๐ฅณ github.com/NVlabs/ProtoMoโฆ Reconstructs (almost) all AMASS motions. Trained using IsaacLab and supports irregular terrains. Next up -- releasing SMPL MaskedMimic compatible with ProtoMotions v2 ๐
00:00
00:00 - AI rendering engines are going to be very fun to play with! Here we use Cosmos to produce videos of a person painting graffiti on the wall. It's really cool to see how the scene structure + motion are maintained, while the "style" is controlled through text prompts.
00:00 - Gym policy running in Lab. ๐ฅณ
00:00
00:08A bit later and some sprinkles of AMP. Now lets get this working in IsaacLab. - MaskedMimic pre-trained model public release ๐งโ๐ github.com/NVlabs/ProtoMoโฆ Some info in the thread on how to play with the model 1/
00:00Excited to share our latest work! ๐คฉ Masked Mimic ๐ฅท: Unified Physics-Based Character Control Through Masked Motion Inpainting Project page: research.nvidia.com/labs/par/maskeโฆ with: Yunrong (Kelly) Guo, @ofirnabati, @GalChechik and @xbpeng4. @SIGGRAPHAsia (ACM TOG). 1/ Read - 1/ Excited to share that our latest work, Conditional Adversarial Latent Models [CALM], has been accepted to @siggraph 2023. ๐งต๐ #reinforcementlearning #animation #games #isaacgym #siggraph2023 @NVIDIAAI
- First day as a research scientist (deep reinforcement learning) at @NVIDIAAI ! Looking forward to the amazing challenges ahead ๐คฉ
- ProtoMotions v2 is out ๐งต github.com/NVlabs/ProtoMoโฆ New and improved scene creation for easily creating consistent scenes across different simulators. Dramatically simplified workflow! Better design with less inheritance. You asked, we delivered: Genesis support now included!
- I got the G1 to cartwheel (in sim). But not to perform a back-handspring yet. Probably a combination of algorithm (reward/termination/ structure and/or L2-tracking vs discriminative) and retargeting not fine-tuned well enough. I'd bet the real G1 can't cartwheel on the static
00:00
00:00 - Just like deep mimic, these methods from "animation" will slowly creep into robotics. First show it works in sim, then handle the sim to real challenges. Amazing results @liangpan_t !This year, an important mission is to let simulated humanoids solve long-term manipulation tasks in challenging dynamic, cluttered environments. This is a simple demonstration achieved by TokenHSI (CVPR 2025). The humanoid here is controlled by a single, unified model.
00:00 - ๐ฅณ Model Release ๐ฅณ Pre-trained MaskedMimic model for the SMPL humanoid released for ProtoMotions v2. Check it out here: github.com/NVlabs/ProtoMoโฆ Video shows a long-horizon inbetweening (target-pose) objective on irregular terrains.
00:00
01:53Playing around with an in-betweening objective. I have no idea how I'd try tackle this directly using RL. Imitating human motions (mask and fill in the blanks) makes this very simple. Goal here is to reach a target pose at timestep T. x.com/ChenTessler/stโฆ - Large codebase refactor. Getting ready for the public release. Looks like I didn't break anything (yet). Woohoo! ๐คฃ
00:00
